Every game now has a two-voice commentary, stitched from a library of voiced phrases
A basketball game on this site now comes with two hosts. One calls the plays, the other answers with the colour: who drove past whom, who got the board, whose foul it was. Every line is about something that really happened in the replay you are watching, and the hosts fill the quiet between calls with a little talk about nothing in particular, the way a broadcast does.
Why a library
Voicing a whole game from scratch takes close to an hour, longer than the game itself lasts. That does not work for a league that plays a game every few minutes of its own time. So the work moved to before the game: a language model writes the phrases for each kind of play, a made three, a block, a foul, a stop at the rim, and for the banter in between. Each phrase is voiced once, and so is every player's name and every team's name. When a game is played, the commentary for it is assembled from those pieces: pick a phrase for the play, drop in the names, join the clips. In our measurements a game's commentary is put together in a couple of minutes, where voicing it whole took close to an hour.
Only what happened
A line may only say what the replay shows. Each one is checked against the plays before it is voiced: a name that is not in the play, or a number the play does not give, and the line is dropped rather than patched. The same rule applies to the library, which a phrase must pass before it is kept: no stray numbers, only the names of the players and teams in that very play.
More of the game called
At first a game got about 150 lines. Measuring one full game showed why: most plays had no line at all, and when two calls came close together the later one was thrown away. Now a defensive rebound, a foul, a drive that was stopped and the stop's defender each get a call, a line that would talk over another waits a few seconds for its turn instead of vanishing, and the banter takes whatever gap is left. In the same full game measured against the library as it stands, that comes to roughly 370 lines instead of about 170.
It will not be perfect
Joined speech can sound a little choppy, and some phrases the model wrote are odd. The pieces are checked by listening to them read back, and the library keeps growing, so the commentary will get better the more games it calls. Open any game's replay and press the sound button to hear it.